Having two variables, an explanatory one ( X o ) and a response one ( Y o ), linked by the classical relation Y^o = g(\bf X^o) + \varepsilon , we want to estimate the function g (.) without any parametric assumption. However, in a lot of situations, the variables are not measured directly but through their proxies \bf X = \bf X^o + \bivarepsilon and Y = Y^o + \eta where and -are the measurements errors. We propose here a new method for estimating the function g (.) in such a context. Our estimator is based on deconvoluted kernels. Uniform convergence is established for strongly mixing stochastic processes. Some simulations show that our estimator is tractable and performs relatively well in practice.